Ericsson: AI Engineer (Campus)

• LSTM cell: a cell state (long-term memory) plus three gates:
• Forget gate: what to drop.
• Input gate: what new information to add.
• Output gate: what to expose as the hidden state.
• Why it helps: the additive cell-state updates let gradients flow over long sequences, which avoids the vanishing gradients of plain RNNs.
• Why I chose it: "Network KPI time series with daily and weekly patterns; the LSTM beat ARIMA by 12% MAPE." Also mention that Temporal CNNs or Transformers are alternatives, and why you didn't use them (data size, latency).

• Baselines: seasonal decomposition (STL) and residual z-scores per cell or site; EWMA control charts.
• ML: Isolation Forest on multivariate KPIs; LSTM/Transformer forecasters, flagging large residuals; autoencoders on KPI vectors.
• Practicalities: a different seasonality per site, alert fatigue (aggregate and correlate alarms), root-cause hints (neighbouring cells, config changes).
• Evaluate with labelled incidents (precision and recall of alerts, time to detect).

• Docker packages an app with its dependencies into an image; containers run it consistently anywhere ("works on my machine" solved).
• vs VM: containers share the host OS kernel, so they are lightweight and start in seconds. VMs virtualise the hardware with a full guest OS, giving stronger isolation but at a heavier cost.
• In ML: reproducible training and serving environments, easy deployment to Kubernetes.

• NOC copilot: summarise alarms and incidents, suggest likely root causes from past tickets (RAG), draft change plans.
• Natural-language intent → configuration: translate operator intents into validated configuration templates (with human approval and simulation).
• Knowledge assistants over 3GPP standards and vendor documents.
• Guardrails: never push config changes without validation; audit trails; strict access control.

Answer (3-minute template):

1. One sentence: "I built an LSTM model that forecasts cell-tower traffic 24 hours ahead to save energy."
2. Data: source, size, preprocessing (two sentences).
3. Model: why LSTM, its key parameters, the baseline comparison.
4. Result: "MAPE 8.4% against 9.6% for ARIMA, enabling sleep mode at off-peak sites."
5. Your role and learnings. - Rehearse aloud and record yourself. Clarity was the deciding factor in this report.

• Slicing: logically separate virtual networks on shared infrastructure, each with guaranteed characteristics (e.g. low-latency for industrial IoT, high-bandwidth for video).
• ML uses: forecasting slice demand, dynamic resource allocation (RL), SLA violation prediction, anomaly detection per slice.